Light Gradient Boosting Machine-Based Fatality Prediction of Single-Vehicle Motorcycle Crashes

In this study, we utilized the light gradient boosting machine (LightGBM) model for prediction analysis of single-vehicle motorcycle crash fatality in Japan. We develop the model using Japan’s traffic crash data from 2019 to 2022. Since the single-vehicle motorcycle crash dataset has an imbalanced distribution property between fatal and non-fatal accidents, it has difficulty predicting fatal cases due to predicting bias on non-fatal cases. To address this issue, we observed the effect of resampling techniques such as Synthetic Minority Oversampling Technique (SMOTE), Edited Nearest Neighbors (ENN), and SMOTE-ENN on the LightGBM prediction results. Furthermore, our LightGBM using Shapley additive explanations (SHAP) identified the most influential features in predicting single-vehicle motorcycle crash fatalities. The experiment based on evaluation metrics shows that combining LightGBM and resampling techniques significantly improved the model performance based on evaluation metrics such as accuracy, AUC, ROC curve and G-mean. This confirms the effectiveness of this approach in reducing the impact of the imbalanced dataset and improving single-vehicle motorcycle crash fatality predictions. Moreover, based on the SHAP value analysis, crash type, road alignment, collision manner, terrain, and motorcyclist age significantly affect the fatality of single-vehicle motorcycle crashes.